Skip to content

Causal AI-driven secure millimeter-wave beam alignment based on a defensive variational autoencoder

Causal AI-Driven Secure mmWave Beam Alignment with Defensive Variational Autoencoders

arXiv: 其他 AI 安全论文· arXiv:2611.18321· Nasir Khan, Asmaa Abdallah, Abdulkadir Celik, Ahmed M. Eltawil, Sinem Coleri, Cicek CavdarSchool of Electrical Engineering and Computer Science, KTH Royal Institute of Technology; Computer, Electrical, and Mathematical Sciences and Engineering Division, King Abdullah University of Science and Technology; School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, U.K; Department of Electrical and Electronics Engineering, Koc UniversityIndexed by this site · 4 days ago· Originally publishedDefenses· Its place in the scan record →

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

The authors propose an mmWave beam alignment framework with causal feature selection plus DVAE defense, claiming scanning overhead is reduced by 89.8% versus exhaustive search and Top-1 accuracy under attack is about 85%.

Threat modelThe attacker injects ℓ2-bounded perturbations into RSSI measurements (bounded by PSR) to cause the beam classifier to misselect. semi-white box: knows the simplified DNN structure and parameters but not the DVAE defense; black-box: trains a surrogate CNN to generate transferable perturbations. The victim system is the BS-side DNN beam classifier.

Questions
In mmWave beam alignment, DL methods learn only correlations, have poor interpretability and high sweep overhead, and are easily misled by adversarial perturbations targeting RSSI measurements, so both efficiency and robustness are needed.
Methods
Use DirectLiNGAM to learn a causal DAG between RSSI and the optimal beam, select a causally relevant feature subset to train a simplified DNN; then use DVAE to reconstruct perturbed inputs into clean representations and fine-tune the classifier end-to-end.
Experiments and results
On the DeepMIMO Boston5G simulation, the authors state that Causal FS reaches 86% Top-1 with 13 sensing beams, with sweep overhead reduced by 89.8% compared with exhaustive; DVAE defense recovers 80–85% Top-1 at PSR≥-20dB, on average exceeding defensive distillation by 15–20%.
Limitations and future work
The authors state that future work needs to adapt to time-varying channels and mobile environments; the method is based on observational causal discovery and does not include do-calculus interventions; DVAE is trained only with semi-white box perturbations; experiments are single-BS, static-scenario simulations without real hardware validation.
Experimental setupSimplified DNN、DVAE · DeepMIMO Boston5G · Top-k Accuracy、Beam Sweep Time et al.
Evaluated models
Simplified DNNDVAE
Benchmarks
DeepMIMO Boston5G
Metrics
Top-k AccuracyBeam Sweep TimeEffective Spectral EfficiencyPSR
AI summaryThe study proposes a causality-aware millimeter-wave beam alignment framework that uses a defensive variational autoencoder (DVAE) to defend against adversarial attacks. Causal beam selection accuracy is comparable to conventional methods, single-sample computational cost is 829 times lower than SHAP, and beam scanning overhead is reduced by 59.4% and 89.8% compared with conventional DL methods and exhaustive search, respectively. DVAE denoises adversarial inputs, maintaining 85% accuracy under different perturbation bounds and averaging 15–20% higher than defensive distillation across attack scenarios.

The study proposes a causality-aware millimeter-wave beam alignment framework that uses a defensive variational autoencoder (DVAE) to defend against adversarial attacks. Causal beam selection accuracy is comparable to conventional methods, single-sample computational cost is 829 times lower than SHAP, and beam scanning overhead is reduced by 59.4% and 89.8% compared with conventional DL methods and exhaustive search, respectively. DVAE denoises adversarial inputs, maintaining 85% accuracy under different perturbation bounds and averaging 15–20% higher than defensive distillation across attack scenarios.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    Addresses the problems in mmWave beam alignment that DL methods have high scanning overhead, lack causal interpretability, and are susceptible to adversarial attacks.

    Core problem: mmWave beam alignment must both reduce scanning overhead and maintain the accuracy and interpretability of DL beam classifiers under adversarial perturbations.

    • Scenario and importance: mmWave/6G relies on narrow beamforming, and initial access requires beam alignment; exhaustive scanning of the full codebook incurs high latency and energy consumption (Introduction).
    • Shortcomings of existing methods: The authors state that existing DL beam alignment only learns statistical correlations between RSSI and the optimal beam, ignoring causal relationships, leading to poor interpretability, weak generalization, redundant scanning, and sensitivity to RSSI perturbations (Abstract, Introduction).
    • Core observation: There is a learnable causal structure between RSSI measurements and the optimal beam index; selecting a small number of causally relevant sensing beams can maintain accuracy, while a VAE can reconstruct perturbed inputs back to clean representations.
    • Proposed approach: two-stage causal beam selection (DirectLiNGAM learns the DAG, then features are selected) + defensive VAE (DVAE) denoising, with end-to-end fine-tuning of the classifier.
    • Threat model:
      • Goal: inject ℓ2-bounded perturbations into RSSI measurements so that the DNN selects the wrong beam, with perturbations kept within the legal power range (Eq. 10).
      • Knowledge: semi-white box knows the simplified DNN structure and parameters but not the DVAE defense; black-box does not know the model and relies on transferable perturbations from a surrogate CNN.
      • Capability: inject perturbations into the RSSI reporting link (over-the-air).
      • Victim system: the BS-side beam predictor based on the simplified DNN.
  2. What related research exists?

    Related work falls into four categories: DL beam alignment, feature selection, causal AI in wireless, and adversarial attack and defense; this paper positions itself as the first to combine causal selection with VAE defense.

    DL beam alignment

    • Ma et al. [8], Heng et al. [9], Yang et al. [10]: use DNNs to predict the optimal narrow beam from partially sensed beams, learn probing codebooks, or perform hierarchical search. The paper notes that these works rely on explicit propagation channel knowledge during training, which is difficult to obtain in dynamic environments.
    • Demirhan et al. [11], Lim et al. [12], Ma et al. [13]: use auxiliary context and historical channels for beam prediction and tracking.

    Feature selection

    • Cousik et al. [6]: model input beam selection as a sequential feature selection problem.
    • Khan et al. [7], [18]: use XAI methods such as SHAP to rank RSSI importance. The paper notes that such methods require iterative selection and repeated retraining, with high computational complexity and poor scalability.

    Causal AI in wireless

    • Roy et al. [14]: causal reinforcement learning for integrated sensing and communication action selection.
    • Thomas et al. [15], Wheeler et al. [16]: causal semantic communication, learning causally invariant representations or using DirectLiNGAM to estimate the causal graph of semantic variables.
    • Karaca [17]: use PC and GES for intent-based network automation. The paper notes that these works focus on semantic modeling, whereas this paper performs model-agnostic input selection for beam alignment.

    Adversarial attack and defense

    • Kim et al. [19]: construct adversarial attacks against beam classification without providing effective countermeasures (as stated in the paper).
    • Catak et al. [22]: adversarial training against white-box/black-box attacks; the paper notes that it needs to cover all attack scenarios and is prone to overfitting.
    • Kuzlu et al. [23]: defensive distillation defense; the paper notes that it is fragile under high perturbation bounds.
    • Sahay et al. [24], Kokalj-Filipovic et al. [25]: denoising autoencoder defense; the paper notes that their attacks are based on standard FGSM, with coarse perturbations and low deception success rates.

    Baselines and benchmarks: beam selection compared with SVD perfect channel, Exhaustive O-DFT, SHAP FS, 2-Tier Hierarchical, Fixed Beam Selection; defense compared with defensive distillation [23]. Simulations use the DeepMIMO Boston5G scenario.

    The authors claim this paper is the first framework in the literature to introduce causal discovery into beam management combined with DVAE defense (Section I-B).

  3. How does the paper address this problem?

    Two-stage framework: DirectLiNGAM causal feature selection for dimensionality reduction, then DVAE denoising and end-to-end fine-tuning of the classifier.

    Overall approach: first, offline, use causal discovery to select a causally relevant subset from RSSI to train a simplified DNN; then, online, use DVAE to reconstruct potentially perturbed inputs into clean representations before feeding them to the classifier.

    Problem setting

    • Downlink mmWave: the BS is equipped with a 32-element ULA, measures RSSI with Mw=32 DFT sensing beams, and selects the optimal beam from the O-DFT codebook (Y=128 narrow beams) (Eq. 5).
    • Dataset D = {(xu, yu)}, where xu is the Mw-dimensional RSSI and yu is the optimal beam label obtained by noiseless exhaustive search; assume yu is not a cause of any RSSI feature (structural constraint).

    Causal beam selection (Algorithm 1)

    • Use DirectLiNGAM to estimate a causal DAG on the joint variables z = [xu, yu]: z = Bz + e, where B is a strictly lower-triangular causal effect matrix and e is non-Gaussian independent disturbance; non-Gaussianity makes the causal direction identifiable.
    • Feature selection: first take the direct parent nodes of the target y (By,i ≠ 0), then fill up by ranking each feature's total connectivity (in-degree + out-degree) until |S| = M_fw.
    • The authors state that this method is observational causal discovery and does not involve do-calculus interventions; under LiNGAM identifiability conditions it recovers a unique causal order, which is stronger than SHAP's marginal contribution explanation (Remark).

    Simplified DNN training

    • Simplified classifier f(·;θ): input M_fw-dimensional RSSI, output Y=128-dimensional beam scores, trained with cross-entropy (Eq. 8).
    • Architecture: three fully connected layers (128/64/32, ReLU), Adam, learning rate 1e-4, 100 epochs, batch 256 (Table II).

    Adversarial attack and DVAE defense (Algorithm 2)

    • Attack: perturbation δu,S = ϵ·gu/‖gu‖2, where gu is the gradient of the loss with respect to the input; semi-white box uses the BS-side DNN gradient, black-box uses the surrogate CNN (three-layer Conv1D) gradient (Eqs. 9, 10).
    • DVAE: the encoder outputs latent variable mean and variance, reparameterized sampling, decoder reconstruction; training input is clean + perturbed samples concatenated, target is clean samples. Loss LDVAE = αℓr + βℓKL, α=0.9, β=0.5, latent dimension 8, 200 epochs, batch 500 (Eq. 11, Table III).
    • End-to-end fine-tuning: jointly optimize L_DVAE + L_CE (Eq. 12) so the classifier adapts to the denoised representation.
    • The paper states: DVAE training uses only semi-white box perturbations; the authors believe that because it is stronger, it can transfer to improve robustness against black-box.
  4. What experiments did the paper conduct?

    Evaluated on the DeepMIMO Boston5G simulation across five aspects: feature selection, beam accuracy, SNR/spectral efficiency, attack and defense, and computational overhead.

    Experimental setup

    • Scenario: DeepMIMO Boston5G, 200m×230m urban area, single BS (32-element ULA, 15m high), single-antenna UE (2m high), 28 GHz, 500 MHz bandwidth, Dtrain 57,144 entries, 70/10/20 split (Table I).
    • Codebook: 32 DFT wide beams for sensing; narrow beams O-DFT ×4, 128 in total.
    • Evaluated models: simplified DNN (three fully connected layers); black-box attack uses a surrogate CNN (three-layer Conv1D); DVAE latent dimension 8.
    • Baselines: SVD perfect channel, Exhaustive O-DFT, SHAP FS, 2-Tier Hierarchical, Fixed Beam Selection; defense baseline defensive distillation.
    • Metrics: Top-k accuracy, sweep time Tsweep, effective spectral efficiency (Eq. 13), PSR (ϵ = √PSR·‖xu‖2).
    • Number of features: validation accuracy saturates at M_fw=13, so all simulations use M_fw=13.

    Main results

    Table IV: Top-k (k=1,2,3) classification accuracy (%), No Attack baseline Top-1 is 90.3.

    Swipe horizontally to read the table

    AttackPSRNo Defense Top-1DVAE(adv) Top-1DVAE(fine-tuned) Top-1Def. Distillation Top-1
    White-box-20 dB79.585.290.288.4
    White-box-10 dB45.583.287.469.9
    White-box0 dB4.772.485.445.3
    Black-box-20 dB81.887.391.590.3
    Black-box-10 dB47.887.589.872.6
    Black-box0 dB7.287.387.947.8
    • The authors state that without defense, Top-1 at PSR=-10 dB and 0 dB drops by more than 40% and 80% relative to the baseline, respectively; DVAE recovers to 80–85% at PSR≥-20dB, and after end-to-end fine-tuning is within 2–3% of the clean baseline, on average exceeding defensive distillation by 15–20%.
    • The authors state that the black-box attack is weaker than the semi-white box because the surrogate and BS model gradient directions do not match.
    • The authors state that DVAE is trained only with semi-white box perturbations yet is also robust to black-box attacks it was not trained on.

    Causal feature selection and beam accuracy

    • Figure 3/4: at M_fw=13, DirectLiNGAM selects 13 features among x10, x13–x24, overlapping 12 with SHAP's top 13.
    • Figure 5: Causal FS reaches 86% Top-1 with 13 sensing beams, rising to 97% Top-2, on par with SHAP FS; Fixed Beam Selection needs 23 beams to reach 86%; 2-Tier Hierarchical performs worst (the authors attribute this to hierarchical error propagation).

    SNR, spectral efficiency and overhead

    • Figure 6: Causal FS with 13/32 beams achieves average SNR close to exhaustive and SVD; the authors state that sweep overhead is reduced by about 89% compared with exhaustive.
    • Figure 7: at M_fw=8, Causal FS spectral efficiency exceeds the SVD scheme and is on par with SHAP FS; exhaustive sweep over 128 beams has the worst spectral efficiency.
    • Sweep time: exhaustive about 10 ms (128 beams), this paper's 13 beams about 1.01 ms, a reduction of about 90%; online inference DNN 2.83×10⁻³ s + DVAE 0.65×10⁻³ s.

    Other ablations and analyses

    • Figure 8: under PSR=-10 dB white-box perturbation, the mean squared error of DVAE-reconstructed RSSI is below 0.1%.
    • Table V: feature selection time, with 4,000 samples SHAP 12,611.96 s, DirectLiNGAM 14.98 s; with 1,000 samples 3,189.08 s versus 7.85 s.
    • The abstract gives: Causal FS per-sample computational cost is 829× lower than SHAP, sweep overhead is reduced by 59.4% compared with traditional DL selection and by 89.8% compared with exhaustive.
    • Complexity: simplified DNN O(L·Hmax²) per sample, DVAE adds O(S·H_DVAE²), H_DVAE=32 ≪ Hmax=128.
  5. What are some points that can be further explored?

    The authors explicitly give only time-varying channel adaptation as a future direction; experiments are static single-BS simulations with limited coverage.

    Limitations and future directions noted by the authors

    • Future work: extend the causal AI framework to dynamically adapt to time-varying channels and mobile environments (Section VII).
    • Method boundary: the method is observational causal discovery and does not consider interventional causality in the do-calculus sense; causal order recovery relies on LiNGAM identifiability conditions (Section IV-B Remark).
    • Environment dependence: causality depends on the propagation environment; when the environment changes significantly, features must be reselected and the DNN retrained/fine-tuned; the authors state that because large-scale characteristics evolve slowly, updates are expected to be infrequent (Section IV-C).
    • Training coverage: DVAE is trained only with perturbations generated under the semi-white box threat model; robustness to black-box comes from transfer (Section V-C, VI-E).

    Scope of experimental coverage

    • The evaluated model is one simplified DNN; the attack surrogate is a three-layer Conv1D surrogate CNN; the only defense comparison is defensive distillation.
    • The simulation is a single-BS, single-antenna UE, static Boston5G scenario, PSR takes three levels -20/-10/0 dB, and the attack is a gradient-normalized perturbation (FGSM-style).
    • Feature selection time comparison is conducted with 1,000–4,000 samples and Mw=32 (Table V).
    • The paper does not report real hardware or measured-channel validation, does not report variance or statistical tests from repeated experiments, and does not evaluate white-box adaptive attacks against the known DVAE defense.
  6. Summarize the main content of the paper

    An mmWave beam alignment framework with causal feature selection plus DVAE defense, balancing low scanning overhead, interpretability and adversarial robustness.

    A defense-oriented paper on 6G mmWave beam alignment, addressing the problems that DL beam classifiers have high scanning overhead, lack causal interpretability, and are easily misled by RSSI adversarial perturbations.

    • Method: Use DirectLiNGAM to learn the causal DAG between RSSI and the optimal beam, take the target's direct parent nodes plus high-connectivity features, select M_fw=13 sensing beams to train a simplified DNN; then use DVAE to reconstruct perturbed inputs into clean representations, and fine-tune the classifier end-to-end with L_DVAE+L_CE.
    • Efficiency: The authors claim Causal FS reduces overhead by 89.8% versus exhaustive scanning and by 59.4% versus conventional DL selection, with per-sample computational cost 829× lower than SHAP; 13-beam scanning takes about 1.01 ms, exhaustive about 10 ms.
    • Accuracy: On DeepMIMO Boston5G, 13 beams achieve 86% Top-1 and 97% Top-2, on par with SHAP FS and close to exhaustive and SVD in SNR (Figure 5, 6).
    • Robustness: Without defense, Top-1 is only 4.7% under white-box PSR=0dB; DVAE (fine-tuned) recovers to 85.4%, exceeding defensive distillation by 15–20% on average across settings (Table IV).
    • Conclusion: The authors claim the framework unifies interpretability, efficiency and security of beam alignment, and future work needs to adapt to time-varying channels and mobile environments; the method is limited to observational causal discovery, and DVAE is trained only with semi-white box perturbations.
Read originalarxiv.org